New computer that doesn’t benchmark as fast as I expected.

[jimbux]

I’d like to get some thoughts from any of you “build your own computer” folks. It’s possible that nothing is amiss with this new computer and my expectations or assumptions are wrong OR it’s time for me to learn a bunch of new stuff so I can make it perform up to snuff!

Basically I’m comparing the time of rendering a ray-traced image (which is a highly parallelizable task - suitable for multi-core procs) on one of my old machines vs the time for the same image on a new machine I just built. I plugged each into the same switch - and they are using the same nas to read the data file from, and write the image to. There are no textures being read and the time to render the overall image should far outweigh read/write times to the nas.

When building both machines I took fairly careful steps to get the CPU paired up with the right motherboard and picked the correct (good quality) memory etc.

Old: AMD Phenom 9850 Quad-core processor. 4 X 2.5 GHz cores
New: AMD Phenom II x6 1090T processor. 6 x 3.2 GHz cores

I expect the new machine to be roughly (6*3.2/4*2.5)=1.92 times as fast as the old one when ray tracing images. On the plus side, the new machine has the new faster SATA connection to the HD (which is a Western Digital 6.0Gb/s SATA drive), plus the memory in the new machine is faster, however some time may be lost splitting up the task to the two extra processors therefore I still expect that the new computer will be (plus or minus a small factor) around 1.9 times as fast as the old one. My render test shows that it is only 1.5 times as fast.

So I ran some other benchmarks using “hardinfo” here on my F11 setup. (Man the PC world has a plethora of benchmarking and testing/monitoring/bios-tweeking tools - I think hardinfo is about all we have right?)

Here’s the results of those tests:

Benchmarks from hardinfo
************************

CPU Blowfish (smaller better)
------------
AMD Phenom 9850 Quad-core     2500 MHz        3.911
AMD Phenom II x6 1090T        3200 MHz        2.792
Speed Ratio: 1.4

CPU CryptoHash (bigger better)
--------------
AMD Phenom 9850 Quad-core     2500 MHz        332.053
AMD Phenom II x6 1090T        3200 MHz        566.166
Speed Ratio: 1.7

CPU Fibonacci (smaller better)
-------------
AMD Phenom 9850 Quad-core     2500 MHz        2.383
AMD Phenom II x6 1090T        3200 MHz        2.493
Speed Ratio: 0.95 (slower?!)

CPU N-Queens (smaller better)
------------
AMD Phenom 9850 Quad-core     2500 MHz        8.998
AMD Phenom II x6 1090T        3200 MHz        0.841
Speed Ratio: 10.7 (wow but huh?!)

FPU FFT (smaller better)
-------
AMD Phenom 9850 Quad-core     2500 MHz        2.056
AMD Phenom II x6 1090T        3200 MHz        1.834
Speed Ratio: 1.12

FPU Raytracing (smaller better)
--------------
AMD Phenom 9850 Quad-core     2500 MHz        10.983
AMD Phenom II x6 1090T        3200 MHz        11.688
Speed Ratio: 0.94 (slower?!)

These don’t make much sense to me. Nothing indicates that my new machine is approximately 2x faster. I would at least expect to see that it’s 28% faster in the case that the test is single threaded (i.e. 3.2/2.5 = 1.28).

Anyway, here I stand trying to figure out what do make of all this and if it's reasonable for me to think I can improve my new machine's performance (w/o overclocking) in the BIOS etc.

Would love any useful insights from you fine folks.

Regards,
James.

[droidhacker]

Your expectations are unrealistic.
Very simply, an increase in the total number of available processor cycles does NOT map linearly to an increase in performance.

SPECIFICALLY, you are probably spending a greater proportion of CPU time WAITING. This happens when it is necessary to load data from ANY slower form of memory into the L1 cache (the only memory level that the CPU does NOT have to wait for).

Also, just because the NAS is able to supply the data *AT A RATE* greater than the processor can consume it does NOT mean that there is ZERO LATENCY when a piece of data is requested. Running at 3.2 GHz, the latency would literally have to be 1/3.2 BILLION SECONDS = about 3 NANOseconds in order to avoid waiting. Good luck with that!

The only way to have a linear increase in performance is for the memory heirarchy to also scale linearly with the processor... both in terms of latency, as well as in terms of THROUGHPUT. Your entire memory heirarchy (which includes your NAS) would have to be TWICE AS FAST and respond in HALF THE TIME in order for you to achieve the performance boost you were expecting.

1) This is why overclocking is just for show.
2) At 1.5x, you did GREAT.

Originally Posted by jimbux

I’d like to get some thoughts from any of you “build your own computer” folks. It’s possible that nothing is amiss with this new computer and my expectations or assumptions are wrong OR it’s time for me to learn a bunch of new stuff so I can make it perform up to snuff!

Basically I’m comparing the time of rendering a ray-traced image (which is a highly parallelizable task - suitable for multi-core procs) on one of my old machines vs the time for the same image on a new machine I just built. I plugged each into the same switch - and they are using the same nas to read the data file from, and write the image to. There are no textures being read and the time to render the overall image should far outweigh read/write times to the nas.

When building both machines I took fairly careful steps to get the CPU paired up with the right motherboard and picked the correct (good quality) memory etc.

Old: AMD Phenom 9850 Quad-core processor. 4 X 2.5 GHz cores
New: AMD Phenom II x6 1090T processor. 6 x 3.2 GHz cores

I expect the new machine to be roughly (6*3.2/4*2.5)=1.92 times as fast as the old one when ray tracing images. On the plus side, the new machine has the new faster SATA connection to the HD (which is a Western Digital 6.0Gb/s SATA drive), plus the memory in the new machine is faster, however some time may be lost splitting up the task to the two extra processors therefore I still expect that the new computer will be (plus or minus a small factor) around 1.9 times as fast as the old one. My render test shows that it is only 1.5 times as fast.

So I ran some other benchmarks using “hardinfo” here on my F11 setup. (Man the PC world has a plethora of benchmarking and testing/monitoring/bios-tweeking tools - I think hardinfo is about all we have right?)

Here’s the results of those tests:

Benchmarks from hardinfo
************************

CPU Blowfish (smaller better)
------------
AMD Phenom 9850 Quad-core     2500 MHz        3.911
AMD Phenom II x6 1090T        3200 MHz        2.792
Speed Ratio: 1.4

CPU CryptoHash (bigger better)
--------------
AMD Phenom 9850 Quad-core     2500 MHz        332.053
AMD Phenom II x6 1090T        3200 MHz        566.166
Speed Ratio: 1.7

CPU Fibonacci (smaller better)
-------------
AMD Phenom 9850 Quad-core     2500 MHz        2.383
AMD Phenom II x6 1090T        3200 MHz        2.493
Speed Ratio: 0.95 (slower?!)

CPU N-Queens (smaller better)
------------
AMD Phenom 9850 Quad-core     2500 MHz        8.998
AMD Phenom II x6 1090T        3200 MHz        0.841
Speed Ratio: 10.7 (wow but huh?!)

FPU FFT (smaller better)
-------
AMD Phenom 9850 Quad-core     2500 MHz        2.056
AMD Phenom II x6 1090T        3200 MHz        1.834
Speed Ratio: 1.12

FPU Raytracing (smaller better)
--------------
AMD Phenom 9850 Quad-core     2500 MHz        10.983
AMD Phenom II x6 1090T        3200 MHz        11.688
Speed Ratio: 0.94 (slower?!)

These don’t make much sense to me. Nothing indicates that my new machine is approximately 2x faster. I would at least expect to see that it’s 28% faster in the case that the test is single threaded (i.e. 3.2/2.5 = 1.28).

Anyway, here I stand trying to figure out what do make of all this and if it's reasonable for me to think I can improve my new machine's performance (w/o overclocking) in the BIOS etc.

Would love any useful insights from you fine folks.

Regards,
James.

[jimbux]

Thanks DH - this is exactly the kind of input I was looking for.

In the case of the ray-tracing render test I did, the time it took to render each image was 10 minutes on the fast machine, 15 on the slower one. I don't believe that the ray-tracer is hitting the disk (nas) during the render apart from when it initially loads the scene into main memory, which I believe is negligable compared to the 10 minute render time.

Still, good to know your thinking regarding the memory and mem-caches etc. I wasn't planning on overclocking the CPU, but when I examine the BIOS settings for the memory frequency (1333 MHz) I believe that the memory I bought (G.Skill Flare F3-16000CL7D-4GBFLS) is rated to be able to run at 1600 MHz or even 2000 MHz. So given your comment about the memory being the bottleneck then the question I have to answer is if it is possible or desireable to run the memory at a fast rate (while not changing the CPU rates etc.). This is new to me - this level of tweeking - so much to learn.

Also - what you say makes sense given a bunch of reviews I've been reading lately comparing the AMD X6 1090T to the Intel i7 980x where the i7 kicks the 1090T's ass, apparently it has much larger mem-caches (amongst other things - also they are $1000 a pop) which the reviewers attribute to the much faster benchmark results from the i7.

Anyway, it's good to know that you think 1.5X is a good improvement - I wouldn't mind squeezing more out if I can - the faster I can crank out frames (images) the better.

Originally Posted by droidhacker

Your expectations are unrealistic.
Very simply, an increase in the total number of available processor cycles does NOT map linearly to an increase in performance.

SPECIFICALLY, you are probably spending a greater proportion of CPU time WAITING. This happens when it is necessary to load data from ANY slower form of memory into the L1 cache (the only memory level that the CPU does NOT have to wait for).

Also, just because the NAS is able to supply the data *AT A RATE* greater than the processor can consume it does NOT mean that there is ZERO LATENCY when a piece of data is requested. Running at 3.2 GHz, the latency would literally have to be 1/3.2 BILLION SECONDS = about 3 NANOseconds in order to avoid waiting. Good luck with that!

The only way to have a linear increase in performance is for the memory heirarchy to also scale linearly with the processor... both in terms of latency, as well as in terms of THROUGHPUT. Your entire memory heirarchy (which includes your NAS) would have to be TWICE AS FAST and respond in HALF THE TIME in order for you to achieve the performance boost you were expecting.

1) This is why overclocking is just for show.
2) At 1.5x, you did GREAT.

[pythagorean]

I built my own computer for the first time not too long ago and I am very happy with the results. So happy that I ended up building two more. I was using Fedora 13 as my OS when I first built the machines but I found myself booting into Windows more often because I wanted to use specific overclocking tools that were only available with in Windows with my motherboards.

So, anyway, what I learned is that the multiplicity of cores leads to a faster overall computer if you are doing specific multitaks that require/use multiple cores.

Instead of building one monstrous 6 core machine I ended up building three different dual core machines that I now have networked. The actual performance of running 3 processes on 3 different dual core machines is better than what you would get from running the same three processes on one 6 core machine. It is cheaper to buy 3 Athlon II x2 7850e Black Edition processors on sale than it is to buy one top of the line 6 core processor.

And I use a KVM switch, too.

[jpollard]

It also depends on what renderer you are using.

If the renderer is not multi-threaded, you will not get the benefit
of the second CPU at all.

[jimbux]

Originally Posted by pythagorean

It is cheaper to buy 3 Athlon II x2 7850e Black Edition processors on sale than it is to buy one top of the line 6 core processor.

I'm surprised that you can make this claim. There are all the other costs associated with buying a CPU, like, mobo, mem, DVDrom reader, case, PSU, cables, etc. Plus the extra electricity to run the machines, and when the number of machines multiplies then you need an AC unit etc. Is it *really* cheaper to run more machines? I'm leaning towards faster chips, fewer computers.

I also build cheaper (more stripped down) machines for non workstation use (render farm) but load up the workstation with goodies for fast interactive use. SSD, kickass graphics card etc...

---------- Post added at 08:47 PM GMT ---------- Previous post was at 08:47 PM GMT ----------

Originally Posted by jpollard

It also depends on what renderer you are using.

Mental ray.

[jpollard]

The claim is reasonable. The top of the line 6 core processor is MUCH more
expensive to make. And I believe it also requires a different mounting
socket - which means a different motherboard - with associated extra
expense.

The lower cost processors (even two) is cheaper - the auxiliary hardware
is also cheaper - so spread some of the savings to a second CPU mount.

[jimbux]

Originally Posted by jpollard

The claim is reasonable. The top of the line 6 core processor is MUCH more
expensive to make. And I believe it also requires a different mounting
socket - which means a different motherboard - with associated extra
expense.

The lower cost processors (even two) is cheaper - the auxiliary hardware
is also cheaper - so spread some of the savings to a second CPU mount.

Perhaps. It's a fairly straightforward argument to settle however. Just have to price out the cost to build 3 X cheaper 2-core machines vs 1 X 6 core and normalize based on render-times. Don't forget to factor in the cost of licences for software for more machines (renderer and render-farm manager software).

[kona0197]

Don't forget to factor in the cost of licences for software for more machines

Actually the cost isn't that bad for the operating system. Microsoft offers a Windows 7 family pack version that can be legally installed on 3 computers. That version is only about $25 more than a stand alone version of Windows 7.

[jimbux]

Originally Posted by kona0197

Actually the cost isn't that bad for the operating system. Microsoft offers a Windows 7 family pack version that can be legally installed on 3 computers. That version is only about $25 more than a stand alone version of Windows 7.

Not to mention the cost for Fedora - good deals to be had there too.

[kona0197]

Agreed. Was just making a point. Sorry.

[jimbux]

Originally Posted by kona0197

Agreed. Was just making a point. Sorry.

I find myself looking for the "like" button (ala facebook) for your post. No sweat Kona - no need to apologize I was just having a bit of fun.

[pythagorean]

To follow up on the costs of three systems vs one 6 core machine.

I bought Biostar AM2+ motherboards for $50 each. I already had powersupplies sitting around from other machines except for I bought one refurbished 500 watt power supply because I wanted to plug a molex cable into the power input for a videocard.

So, there is $180 for motherboards and power supply. The Athalon II x2 7850e chips were $50 each on sale. So, we are up $330. The laptop and desktop I already own came with Windows 7 keys, so I bought one OEM disk and installed Windows 7 on two machines. The third machine dual boots Fedora and Linux Mint. This cost $100. So, now we are up to $430. I bought 3 hard drives at Tigerdirect: 80 GB SATA drives for $35 dollars each. This brings the total to $535.

The biggest expense was RAM. I have 2 GB in each machine for a total of 6GB. Each stick was about $40. The total amount, therefore, was less than $750.

As for the electricity aspect of it all: I can turn on each machine progressively as I need it for real tasks. It isn't like I have to run all three machines all the time. This also allows me to plug the unused RAM from one machine into the one I am using at the time if need be.

Finally, I believe in modularity. I can swap out a motherboard and a processor with upgraded hardware to dramatically improve my overall system performance much cheaper than someone who has a single motherboard with one 6 core processor could do. The same is true of RAM and any of my other components.

So, even if I have in some sense paid a little more than an single machine with 6 cores would have cost, I also have much more flexibility and in the end my overall set up can do more faster.

[jpollard]

Yes.

It is also the same with systems using a backplane architecture instead
of a motherboard.

upgrade everything separately. Even peripheral bus adapters
(northbridge/southbridge) and clock speeds.

Too bad it isn't used much anymore - though it does exist (sort of)
in blade servers.

[jimbux]

Originally Posted by pythagorean

To follow up on the costs of three systems vs one 6 core machine.

<snip>

So, even if I have in some sense paid a little more than an single machine with 6 cores would have cost, I also have much more flexibility and in the end my overall set up can do more faster.

That's really cool that you could economize and build a bunch of systems to meet your needs that was cheaper than had you built one 6 proc machine. Kudos to you for recycling.

But in the general case I don't think your claim is true. For the sake of argument let's not assume people have spare parts laying around, like PSUs and cases (which you neglected to mention + plus the cost of your KVM). Also if the 6 (total) procs need to share data then in your case you need a file server as well, and a switch, where the guy who builds a 6 proc machine (assuming he's happy with exactly six procs and needs no more) can do it all on one machine without the nas and switch.

So for me: I don't know why people feel the need for multiprocs, but I need them for 3d rendering (raytraced) animations - So I need as many procs as I can get my hands on. Each machine has a fairly big cost, all the parts needed to make it, shelf-space, more cables, AC unit to keep the room cool when all the machines are running full tilt, the time it take to be IT-guy for multiple machines (installs, upgrades, etc have to happen on each machine - keeping them in sync as much as possible), possible addition of kvm switch. When I also factor in the cost of software licences (they are per machine not per cpu core) - less machines are way cheaper. And I have to buy all my parts - as I don't have that many spare parts laying around.

And on more point - time is money. When I'm working on lighting/comping a shot, I need my workstation to be as fast as possible to make me as productive as possible. Multi core speeds me up a lot for what I'm doing, there would be no way I could do that by splitting up between multiple machines.

Interesting discussion - thanks.